NONLINEAR AND STRUCTURE-AWARE TEMPORAL MODELING FOR EEG CLASSIFICATION: A MULTI-DOMAIN DEEP LEARNING APPROACH

EEG signals are characterized by strong nonlinearity, complex temporal dynamics, and spatial structures reflecting brain-region organization, posing significant challenges for effective classification. In this work, we propose a multi-domain deep learning method to improve EEG classification performance by combining temporal attention processes, nonlinear modeling, and structure-aware spatial representation. The suggested TEGNet system uses a Transformer-based module to record long-range temporal dependencies, a Graph Neural Network (GNN) to encode spatial topological features among EEG electrodes, and a Kolmogorov–Arnold Network (KAN) to describe nonlinear inter-channel interactions. Using a multi-domain fusion approach, these complementary modules are combined to create a single feature representation. In multi-class and cross-subject classification contexts, experimental results on the BCI Competition IV-2a dataset show that our method performs much better than state-of-the-art baselines. Ablation studies and visualization analyses further verify the individual contributions and interpretability of each module. This work highlights the importance of jointly modeling nonlinear, temporal, and structural features for robust and generalizable EEG decoding.